Soltani et al. (2025) Enhancing Flood Forecasting with Machine Learning Informed by Integrated ParFlow-CLM Hydrological Modeling

Identification

• Journal: Earth Systems and Environment
• Year: 2025
• Date: 2025-11-18
• Authors: Samira Sadat Soltani, Z Mohammadi, Ardalan Izadi, Stefan Kollet
• DOI: 10.1007/s41748-025-00923-5

Research Groups

• Forschungszentrum Jülich GmbH, Institute of Bio- and Geosciences (IBG-3, Agrosphäre), Jülich, Germany
• Centre for High-Performance Scientific Computing in Terrestrial Systems, Geoverbund ABC/J, Jülich, Germany
• Faculty of Civil Engineering, K. N. Toosi University of Technology, Tehran, Iran

Short Summary

This study integrates a fully coupled hydrological model (ParFlow/CLM) with a Gated Recurrent Unit (GRU) Convolutional machine learning model to enhance flood forecasting. It demonstrates that incorporating physically-derived soil water content (SWC) significantly improves the accuracy of river discharge predictions, outperforming standalone AI and hydrological models.

Objective

• To highlight the synergy between machine learning and physics-based modeling, demonstrating how their integration can significantly enhance the reliability of flood forecasting systems.
• To examine the effectiveness of ParFlow/CLM in modeling flood events by assessing its ability to represent hydrological and land surface processes accurately.
• To explore the strengths and limitations of ParFlow/CLM relative to AI-based approaches, evaluating their accuracy, efficiency, and applicability in flood modeling.
• To investigate whether combining AI techniques with ParFlow/CLM can improve model performance, assessing whether these approaches can complement each other for more accurate flood prediction and risk assessment.

Study Configuration

• Spatial Scale: Eifel-Ardennes mid-mountain region (western Germany, eastern Luxembourg, southeastern Belgium), exceeding 20,000 square kilometers. River basins studied include Ahr (749.0 km²), Kyll (175.6 km² to 816.3 km²), and Prüm (53.2 km² to 576.1 km²). Model simulations were performed at a horizontal resolution of 0.0055° × 0.0055° (approximately 0.611 km).
• Temporal Scale: The study focused on the July 2021 extreme flood event, specifically from July 14–15, 2021. ParFlow/CLM simulations were conducted at an hourly time step. The machine learning model incorporated a look-back window of 15 time steps.

Methodology and Data

• Models used:
  • Integrated Hydrological Model (IHM): ParFlow/CLM (ParFlow for three-dimensional, variably saturated subsurface and surface water flow; CLM for land surface processes).
  • Machine Learning Model: Gated Recurrent Unit (GRU) Convolutional model (a deep learning architecture integrating Convolutional Neural Network (CNN) for spatial feature extraction and GRU for temporal dynamics).
• Data sources:
  • Land Surface Data: Corine Land Cover (CLC2018 v20) at 100-meter resolution (reclassified to 18 IGBP land cover types). SoilGrids250m dataset for soil texture.
  • Atmospheric Forcing: High-Resolution (HRES) deterministic medium-range forecast from the European Centre for Medium-Range Weather Forecasts (ECMWF) at 0.1° × 0.1° (downscaled to 0.0055° × 0.0055°). Variables included longwave and shortwave radiation, total precipitation, air temperature, atmospheric pressure, specific humidity, and wind velocity components.
  • Evaluation Data: Water level (W) and streamflow (Q) data from multiple river gauges (e.g., Altenahr-Ahr, Jünkerath-Kyll, Kordel-Kyll, Prüm2-Prüm, Prümzurlay-Prüm) provided by relevant water management authorities (e.g., federal state of Rhineland-Palatinate and Wupperverband).

Main Results

• The integration of ParFlow/CLM-derived Soil Water Content (SWC) as an input feature significantly improved the predictive accuracy of the AI model for flood forecasting.
• At Kordel-Kyll, the AI model with SWC achieved a Root Mean Square Error (RMSE) reduction of up to 51% and an increase in the coefficient of determination (R²) from 0.634 to 0.916, compared to the AI model without SWC.
• The AI model incorporating SWC demonstrated superior performance in capturing both the peak and recession limbs of flood hydrographs, and accurately reproduced the timing of peak discharge at four out of five stations, outperforming both standalone AI and physics-based ParFlow/CLM models.
• Probabilistic assessment using the First-Order Reliability Method (FORM) showed that the AI model with SWC consistently had the lowest probability of failure (Pf), indicating the highest reliability in discharge prediction.
• The benefits of including SWC were found to be basin-dependent, with the most significant improvements observed in flatter, slower-responding catchments where soil moisture plays a dominant role in runoff generation.

Contributions

• Development and validation of an integrated modeling framework that synergistically combines a fully coupled surface-subsurface hydrological model (ParFlow/CLM) with a deep learning architecture (GRU-Convolutional model) for enhanced flood forecasting.
• Quantification of the substantial improvement in flood prediction accuracy achieved by explicitly incorporating physically-derived soil water content (SWC) into a machine learning framework.
• Demonstration of the value of physics-based models as a feature engineering step for machine learning, enriching the input space with hydrologically meaningful variables.
• Provision of a robust and more reliable framework for real-time flood prediction, offering valuable insights for water resource managers and policymakers to mitigate flood-related damages and enhance early warning systems.

Funding

• German Research Foundation (DFG)
• Earth System Modelling Project (ESM)
• Project number: 320397309

Citation

@article{Soltani2025Enhancing,
  author = {Soltani, Samira Sadat and Mohammadi, Z and Izadi, Ardalan and Kollet, Stefan},
  title = {Enhancing Flood Forecasting with Machine Learning Informed by Integrated ParFlow-CLM Hydrological Modeling},
  journal = {Earth Systems and Environment},
  year = {2025},
  doi = {10.1007/s41748-025-00923-5},
  url = {https://doi.org/10.1007/s41748-025-00923-5}
}